Simulation engine validation: 15 checks against queueing theory
The Simflowly engine was run against 15 results that queueing theory gives exactly. The largest relative error was 2.05%, and the exact value fell inside the 95% confidence interval in 13 of 15 comparisons. The two that missed are explained below, with the seeds you can rerun.
What was checked
Queueing theory has a few systems with exact, closed-form answers. We built each one as a model in the Simflowly engine, ran it, and compared the simulated mean with the exact value. Five systems are covered:
- M/M/1 at 80% utilization: random arrivals, exponential service, one server (5 comparisons).
- M/M/3 at 80%: three servers sharing one line, with the Erlang C probability of waiting (5 comparisons).
- M/M/6 at 90%: six servers, a harder case because the line is long (3 comparisons).
- M/D/1 at 80%: constant service time, the Pollaczek-Khinchine result (1 comparison).
- Two M/M/1 queues in series, 80% each: the time in the system from Burke's theorem (1 comparison).
The 15 comparisons
Seed 2026, 30 replications of 100,000 simulated minutes each, 5,000-minute warm-up. "In CI" means the exact value lies inside the 95% Student-t confidence interval of the simulated mean.
| Scenario | Metric | Exact | Simulated mean | 95% interval | Error | Inside |
|---|---|---|---|---|---|---|
| M/M/1, lambda=0.8/min, mu=1/min (rho 0.8) | utilization | 0.8000 | 0.8013 | 0.7998 to 0.8029 | 0.17% | yes |
| M/M/1, lambda=0.8/min, mu=1/min (rho 0.8) | mean wait Wq (min) | 4.0000 | 4.0415 | 3.9707 to 4.1123 | 1.04% | yes |
| M/M/1, lambda=0.8/min, mu=1/min (rho 0.8) | mean queue Lq | 3.2000 | 3.2348 | 3.1758 to 3.2939 | 1.09% | yes |
| M/M/1, lambda=0.8/min, mu=1/min (rho 0.8) | time in system W (min) | 5.0000 | 5.0428 | 4.9713 to 5.1143 | 0.86% | yes |
| M/M/1, lambda=0.8/min, mu=1/min (rho 0.8) | throughput (per hour) | 48.0000 | 48.0206 | 47.9563 to 48.0849 | 0.04% | yes |
| M/M/3, lambda=2.4/min, mu=1/min per server (rho 0.8) | utilization | 0.8000 | 0.8007 | 0.7997 to 0.8018 | 0.09% | yes |
| M/M/3, lambda=2.4/min, mu=1/min per server (rho 0.8) | mean wait Wq (min) | 1.0787 | 1.0850 | 1.0723 to 1.0978 | 0.59% | yes |
| M/M/3, lambda=2.4/min, mu=1/min per server (rho 0.8) | mean queue Lq | 2.5888 | 2.6052 | 2.5732 to 2.6372 | 0.63% | yes |
| M/M/3, lambda=2.4/min, mu=1/min per server (rho 0.8) | P(wait>0), Erlang C | 0.6472 | 0.6480 | 0.6462 to 0.6498 | 0.12% | yes |
| M/M/3, lambda=2.4/min, mu=1/min per server (rho 0.8) | time in system W (min) | 2.0787 | 2.0856 | 2.0724 to 2.0988 | 0.33% | yes |
| M/M/6, lambda=5.4/min, mu=1/min per server (rho 0.9) | utilization | 0.9000 | 0.9005 | 0.8998 to 0.9012 | 0.06% | yes |
| M/M/6, lambda=5.4/min, mu=1/min per server (rho 0.9) | mean wait Wq (min) | 1.2335 | 1.2584 | 1.2411 to 1.2757 | 2.02% | no |
| M/M/6, lambda=5.4/min, mu=1/min per server (rho 0.9) | mean queue Lq | 6.6611 | 6.7975 | 6.7028 to 6.8922 | 2.05% | no |
| M/D/1 (Pollaczek-Khinchine), rho 0.8 | mean wait Wq (min) | 2.0000 | 2.0044 | 1.9823 to 2.0264 | 0.22% | yes |
| Two M/M/1 in series (Burke), rho 0.8 each | time in system W (min) | 10.0000 | 10.0585 | 9.9441 to 10.1729 | 0.59% | yes |
Worst relative error: 2.05%. The exact value is inside the 95% interval in 13 of 15 comparisons. Across 15 comparisons a perfect engine would land inside the interval in about 14 of them on average (15 × 0.95). The comparisons are not independent, because several metrics come from the same run.
The two misses, explained
Both misses are the M/M/6 queue at 90% utilization: the average wait and the average line length. With seed 2026 they came out 2.0% above theory and just outside their 95% interval.
Re-running that scenario with seeds 1 to 4 put the exact value inside the interval every time. The simulated mean waits were 1.2196, 1.2337, 1.2207 and 1.2421 minutes, against the exact 1.2335. Some seeds land above and some below, so this is sampling noise at 95% coverage, not a bias in the engine. The largest error among the other 13 comparisons is 1.1%.
| Seed | Simulated mean wait (min) | Exact (min) |
|---|---|---|
| 1 | 1.2196 | 1.2335 |
| 2 | 1.2337 | 1.2335 |
| 3 | 1.2207 | 1.2335 |
| 4 | 1.2421 | 1.2335 |
Definitions used on every Simflowly surface
- Wq: minutes in line before service starts.
- Lq: time-average number of units waiting (not being served); for the whole system, summed over stations.
- P(wait > 0): share of units that waited before service.
- For M/M/c they satisfy Lq = P(wait > 0) × ρ ÷ (1 − ρ) and Wq = Lq ÷ λ.
What this validation does not cover
Only the five systems above are checked against exact answers: exponential or constant service times, one to six servers, one station or two in series. Other distributions, batching, breakdowns, shared crews, limited queues and routing are features of the engine that this table does not test against theory. The calculators and guides that use them say so, and the template pages show real runs without an exact answer to compare with. Validation also says nothing about your inputs: wrong cycle times give a wrong answer from a correct engine.
Watch the engine run a queueSee the M/M/3 rows worked by hand
Questions
How accurate is the Simflowly simulation engine?
Across 15 comparisons with closed-form queueing results (M/M/1, M/M/3, M/M/6, M/D/1 and two M/M/1 queues in series), the largest relative error was 2.05%. The theoretical value fell inside the 95% confidence interval in 13 of 15 cases. Each result is the mean of 30 replications of 100,000 simulated minutes, with seed 2026.
Why did two comparisons miss the 95% interval?
Both belong to the M/M/6 queue at 90% utilization, where the wait and the line length came out 2.0% above theory with seed 2026. Re-running with seeds 1 to 4 put the theoretical wait inside the interval every time, with means of 1.2196, 1.2337, 1.2207 and 1.2421 against 1.2335, so it looks like sampling noise, not bias.
Does this prove that every model is right?
No. It checks the engine against queues that have exact answers: exponential or constant service, one to six servers, and two stations in series. It does not validate your inputs, and it does not cover every feature, such as batching, breakdowns or shared crews. Your own data and a check against your real process are still needed.